Why AI Cites You in One Answer and Forgets You in the Next

I first wrote about this framework, Total Graph Authority, on the Search Agency blog. This is a shorter version for founders and small marketing teams.

The same prompt gives you a different answer on Tuesday than it did on Monday. Not because your product changed, but because the model reassembled its answer from a slightly different set of sources the second time. Run a buyer question through ChatGPT three times and watch the citations shuffle. That instability is the thing most brands haven't planned for.

It also explains a pattern I see constantly in the prompt sets I track. A brand shows up in one answer, vanishes from the next, and nobody can say why. The honest reason is usually that the brand did one channel well and left the others empty, so it only wins the answers that happen to lean on the channel it owns.

One good channel is a coin flip

Picture a made-up bookkeeping app for Indonesian small businesses called Rapiku. Its site is excellent. Clean pages, real depth, fast to load. Ask an assistant "what's the best bookkeeping app for a small business in Indonesia" and here's roughly what happens across repeated runs.

RUN 1  cites rapiku.example, a Google review page       -> Rapiku named
RUN 2  cites a Reddit thread, a "top 10" listicle        -> Rapiku absent
RUN 3  cites a YouTube walkthrough, a competitor blog     -> Rapiku absent

Rapiku's beautiful site wins the run that leaned on its site. The other two pulled from Reddit, a listicle, and YouTube, and Rapiku had built nothing there. Two answers out of three went to a competitor, and no amount of extra work on the site would have changed that. The gap wasn't quality. It was absence.

Total Graph Authority is the fix. You reach it when three independent kinds of source all point at you and describe you the same way. Your own domain. Other people's sites. Social platforms. When all three agree, it stops mattering which path the model takes, because every path lands on you.

Your own site has to be quotable, not just good

First-party content is where a model learns what you are and lifts clean answers. Depth alone isn't enough. A model has to be able to take a passage whole, without dragging the rest of the page along.

That's a writing decision more than a technical one. Compare two ways of stating the same fact.

WEAK    "Pricing depends on a few things and can get complicated
         depending on your business."

STRONG  "Rapiku has three plans. Free for under 50 invoices a month,
         Pro at IDR 99,000, and Business at IDR 249,000. Only Business
         includes multi-user access and tax-report export."

The strong block gets quoted because it answers the question on its own. The weak one gets skipped because there's nothing in it to lift. Write your key pages in blocks like the second one, then let schema tell the model who you are. An Organization block with a full sameAs array earns its place before any other markup, because it links the Rapiku on your site to the Rapiku on LinkedIn, Crunchbase and Wikidata, and stops your signals from scattering across what looks like three different companies.

Other people vouching for you counts more than your own claims

Here's the part that surprises founders. In the prompt sets I track, third-party pages get cited more often than the brand's own site. A model trusts corroboration from places it doesn't think you control, which means the page you don't own can outrank the one you do.

Most buyer questions in AI are comparative anyway. "Best X", "X versus Y", "alternatives to Z". Those answers come from listicles, directories and forum threads, not vendor pages. You can't publish the neutral "top 10" yourself, but you can earn a spot in the ones people already read, and you can hand reviewers an honest comparison they'll actually use.

A table that admits where a rival wins gets published. One that pretends you win every row gets ignored. A few surfaces get quoted out of all proportion to their size, and those are the ones worth real effort. G2 and Capterra profiles. Reddit and local Facebook groups. A Wikidata entry if you clear the notability bar. And one original data study that gives journalists a number only you have.

Social is a citation surface now, not just reach

Assistants pull from YouTube, TikTok and Instagram for a specific class of question, the "does it actually work for me" ones. "Is Rapiku good for a warung?" "Can it handle PPN reports?" Transcripts and captions get indexed and quoted, and a model reads your title and description before it processes a second of video. So a walkthrough titled "Is Rapiku Worth It for a Small Warung? Honest Review" with a clear "best for small retail doing 20 to 200 invoices a month" line in the description does more for citations than the footage itself.

What to fix first

Order matters, because each layer makes the next one land harder. This is the sequence I run with clients.

1. Fix the entity. One brand name, one canonical site, Organization schema with a complete sameAs, a Wikidata item if you qualify. Everything downstream credits back here.
2. Make your top pages quotable. Turn your money pages into self-contained answer blocks, then add an llms.txt pointing crawlers at them.
3. Earn corroboration. Claim your directory profiles, publish one honest comparison, run one small original study, answer real questions in the forums your buyers use.
4. Add social proof of use. A few eligibility and how-to videos with query-shaped titles.
5. Measure and compound. Track what turns into citations and put more behind it.

Knowing whether it's working

AI visibility is measurable, so treat it like a number, not a vibe. Pick a small, stable set of prompts grouped by intent (branded, category, comparison, problem, eligibility) and run them weekly across ChatGPT, Gemini, Perplexity and Google AI Overviews. Watch four things. How often you're named at all, how much of the cited set is yours or vouches for you, whether the mentions frame you well, and how much traffic actually starts on an AI surface.

A single citation was never the goal. Presence everywhere the model might look is, because that's the only position that survives an answer the model keeps rebuilding.

If you want to see where your brand stands across those three sources right now, the AI Search & SEO Audit benchmarks your visibility against your own prompts and your named competitors.